AgapHost/ai/model-registry.yaml

# Model registry — models are the scarce queued resource.
#
# Per DESIGN-a2a-agents.md v2.1 §2-3b (commit df2071d5), kanboard task #133
# (A2A-1). Version-controlled here; the "model plane" (§3) and the fabric's
# workers/routers read this data — they do not duplicate it. Lifecycle a(t)
# *probe mechanics* (QuotaProbe, GPUResidencyProbe, ...) live in
# kanboard/bin/kb_worker.py; this registry supplies the *parameters* those
# probes consume (commands, fields, thresholds, VRAM footprints).
#
# Scope constraint (alvis, §3a): NO METERED API BY DEFAULT. The workflow is
# Claude Code (a flat-subscription runtime -> agent registry #134, not here)
# + the Kimi wrapper + a local GPU embedder + a small weak local model. The
# governor arbitrates quota and GPU, not money. Any metered model below is
# `metered: true, opt_in_required: true` and carries no default route to it
# (see routing.metered_opt_in: [] at the bottom — empty means unreachable).
#
# Read with model_registry.py (same directory): resolve(), preload_check().
#
# ── Coverage vs litellm-config.yaml (kb#195, 2026-07-26 audit) ──────────
# Every model_name litellm-config.yaml defines must appear either as a
# `litellm_model_name` below or in this exclusion list. litellm_key_spec()
# default-denies anything not reachable via routing.tiers, so an excluded
# model stays ungoverned-but-inert until someone wires it up (add it here
# and to routing.tiers first).
#
#   GOVERNED (present below):
#     - ollama/gemma3:4b -> id: local-small (hot path: Hindsight LLM/
#       consolidation/reflect all route here as of 2026-07-26)
#     - judge -> id: paid-fallback (metered; see kb#164 for the fact that
#       the no-metered-API constraint has no runtime enforcement yet)
#     - codex-agent -> id: codex-agent (LiteLLM-routed large deployment; see
#       below. Was kimi-agent + its own container until the 2026-08-01 purge)
#     - bge-m3 -> id: bge-m3 (kb#164, 2026-07-26: wired into litellm-config
#       .yaml pointing at ollama on 11436, the real embedder/routing
#       classifier; litellm_model_name below updated from null to "bge-m3")
#
#   INTENTIONAL EXCLUSIONS (not governed by this registry, by design):
#     - tip-generator (ollama/qwen2.5:1.5b), embedder (ollama/nomic-embed-
#       text): aliases consumed by the separate oO ml/serving project, not
#       the a2a fabric. Tracked in oO/CLAUDE.md, not duplicated here.
#     - Raw ollama/* passthrough exposures — ollama/qwen3.5:4b,
#       ollama/qwen3:8b, ollama/qwen2.5:1.5b, ollama/qwen2.5:0.5b,
#       ollama/gemma3:1b, ollama/nomic-embed-text — manual/dev-console
#       access to the ollama instances for ad-hoc testing. No agent or
#       fabric workflow is registered against them (grepped agent-registry
#       .yaml and openai/*.py: no hits). Not in routing.tiers, so
#       litellm_key_spec() grants no agent access to them either way.
#       If one of these becomes a real dependency (as ollama/gemma3:4b
#       did), give it its own registry entry at that point.
#     - The 12 OpenRouter `*:free` models (meta-llama/llama-3.3-70b-
#       instruct:free, meta-llama/llama-3.2-3b-instruct:free, deepseek/
#       deepseek-r1:free, qwen/qwen3-4b:free, qwen/qwen3-coder:free,
#       google/gemma-3-27b-it:free, google/gemma-3-12b-it:free, mistralai/
#       mistral-small-3.1-24b-instruct:free, nvidia/nemotron-3-super-
#       120b-a12b:free, openai/gpt-oss-120b:free, minimax/minimax-m2.5:free,
#       nousresearch/hermes-3-llama-3.1-405b:free) — human-facing manual-
#       selection models (e.g.
#       via Open WebUI), outside the agent fabric's model plane. Not
#       referenced by any agent registry entry, not in routing.tiers, so
#       resolve()/litellm_key_spec() never route an agent to them. Free
#       tier, so this is not the kb#164 metered-enforcement gap — flag
#       for a proper entry only if an agent workflow starts depending on
#       one of these.

schema_version: 1

models:
  # ── codex — main reasoning ─────────────────────────────────────────────
  # Flat ChatGPT subscription via `codex login`, wrapped by two independent
  # Codex-CLI containers (own creds volume each). Not behind LiteLLM today —
  # callers hit the wrapper HTTP endpoints directly.
  #
  # Migrated from Kimi CLI 2026-07-31 for cost: this retires the separate
  # Moonshot subscription in favour of the already-paid ChatGPT plan. The id
  # changed `kimi` -> `codex`; routing.tiers below refers to it by id.
  - id: codex
    role: "main reasoning (adolf-llm / hindsight-llm Codex-CLI wrappers)"
    litellm_model_name: null
    endpoints:
      - name: adolf-llm
        purpose: "Adolf's conversational backbone"
        url: "http://adolf-llm:8010"
        # Route retained but returns HTTP 501 since the Codex migration —
        # no machine-readable quota on this backend. See quota: below.
        usage_url: "http://localhost:8010/usage"
      - name: hindsight-llm
        purpose: "Hindsight's structured-extraction LLM (HINDSIGHT_API_LLM_MODEL)"
        url: "http://hindsight-llm:8012/v1"
        model_name: "openai/hindsight-llm"
    tier: large
    context_tokens: 400000  # GPT-5-Codex context window; re-verify if the CLI's pinned model changes
    tool_use_quality: high
    lifecycle: quota-gated
    quota:
      # UNRESOLVED at the Codex migration (2026-07-31): the Kimi backend had a
      # machine-readable managed-usage API that adolf-llm's /usage route
      # normalized into these fields. Codex exposes no equivalent endpoint, so
      # /usage now returns HTTP 501 and there is currently NO quota probe for
      # this model. Governor quota-gating on `codex` is therefore blind — it
      # will not see the ChatGPT plan's rate limits until a signal is found.
      probe_command: null
      windows: []
      threshold_pct: 95
    gpu_residency: null
    cost_class: subscription   # flat-rate, not metered — quota is the constraint, not spend
    metered: false
    opt_in_required: false

  # ── codex-agent — the LiteLLM-routed large deployment ──────────────────
  # Replaces the retired `kimi-agent` container (2026-08-01 Kimi purge). That
  # container was the ONLY large-tier deployment behind LiteLLM, backing
  # `tier-large`, the auto_router's complex-reasoning route and their
  # fallbacks — removing it without a replacement would have silently degraded
  # every large-tier request to the local 4B model. Rather than stand up a
  # third CLI container with its own login, this now points at the existing
  # codex-backed adolf-llm wrapper (:8010, model id "adolf").
  #
  # Same underlying ChatGPT subscription as `codex` above — the two ids differ
  # only in call path (this one via LiteLLM, `codex` direct to the wrappers),
  # so their quota is shared and neither has a probe.
  - id: codex-agent
    role: "LiteLLM-routed large deployment; proxies to the codex-backed adolf-llm wrapper"
    litellm_model_name: "codex-agent"   # openai/litellm-config.yaml model_list entry
    endpoints:
      - name: adolf-llm
        url: "http://adolf-llm:8010/v1"
    tier: large
    context_tokens: 400000  # GPT-5-Codex context window; re-verify if the CLI's pinned model changes
    tool_use_quality: high
    lifecycle: quota-gated
    quota:
      probe_command: null   # no machine-readable quota on the Codex backend — see `codex`
      windows: []
      threshold_pct: null
    gpu_residency: null
    cost_class: subscription
    metered: false
    opt_in_required: false

  # ── local-small — the cheap tier ───────────────────────────────────────
  # ollama/gemma3:4b on the GPU ollama instance. Already the live model for
  # Hindsight consolidation/reflect (HINDSIGHT_API_CONSOLIDATION_LLM_MODEL /
  # HINDSIGHT_API_REFLECT_LLM_MODEL, kb#88) and exposed via LiteLLM.
  - id: local-small
    role: "cheap tier — ollama small/weak local model (background extraction, consolidation, reflect)"
    litellm_model_name: "ollama/gemma3:4b"   # openai/litellm-config.yaml model_list entry
    endpoints:
      - name: ollama-direct
        url: "http://host.docker.internal:11436"
      - name: via-litellm
        url: "http://litellm:4000/v1"
    tier: small
    context_tokens: 8192   # gemma3:4b default ctx; re-verify with `ollama show gemma3:4b` if raised
    tool_use_quality: low
    lifecycle: always-on
    quota: null
    gpu_residency:
      vram_mb: 4000              # approx measured footprint, within the shared 8GB card (see gpu_residency_policy below)
      never_evict: false         # evictable — a bigger model may push it out; that's a silent regression to catch, not prevent here
      co_residency_group: interactive-local
    cost_class: free
    metered: false
    opt_in_required: false

  # ── bge-m3 — embedder + routing classifier ─────────────────────────────
  # Never-evict: it's both Hindsight's recall embedder AND (design §3a) the
  # embedding model LiteLLM Auto Router's semantic-router classifier will
  # use for tier/complexity routing — losing it degrades both recall AND
  # routing at once.
  - id: bge-m3
    role: "embedder — also the routing classifier (§3a, LiteLLM Auto Router / semantic-router)"
    litellm_model_name: "bge-m3"   # kb#164, 2026-07-26: wired into litellm-config.yaml (ollama/bge-m3 @ 11436) -- was null (unwired gap)
    endpoints:
      - name: ollama-direct
        url: "http://host.docker.internal:11436"
        openai_compatible_path: "/v1/embeddings"
    tier: small
    context_tokens: 8192
    tool_use_quality: "n/a"   # embedder, not a chat/tool-use model
    lifecycle: always-on
    quota: null
    gpu_residency:
      vram_mb: 1200
      never_evict: true
      co_residency_group: interactive-local
    cost_class: free
    metered: false
    opt_in_required: false

  # ── tei-reranker — interactive-critical, never-evict ───────────────────
  # Not an LLM (cross-encoder rerank sidecar for Hindsight recall, kb#87)
  # but carries the same GPU-residency stakes as bge-m3, so it's tracked
  # here rather than invented as a separate registry class.
  - id: tei-reranker
    role: "cross-encoder reranker sidecar for Hindsight recall (interactive-critical)"
    litellm_model_name: null   # TEI-compatible /rerank API; not routed through LiteLLM
    endpoints:
      - name: tei-reranker
        url: "http://tei-reranker:80"   # host-published :8014
    tier: small
    context_tokens: null
    tool_use_quality: "n/a"
    lifecycle: always-on
    quota: null
    gpu_residency:
      vram_mb: 1000
      never_evict: true
      co_residency_group: interactive-local
    cost_class: free
    metered: false
    opt_in_required: false

  # ── paid-fallback — optional, opt-in only ──────────────────────────────
  # §3a: "Any paid deployment in the LiteLLM config must be explicitly
  # enabled per agent via its virtual key; nothing routes to a metered
  # model implicitly." routing.metered_opt_in below is the enforcement
  # point: empty list = no caller has opted in = unreachable by resolve().
  - id: paid-fallback
    role: "optional metered fallback (e.g. Haiku) — disabled by default"
    litellm_model_name: "judge"   # litellm-config.yaml's existing entry (anthropic/claude-haiku-4-5-20251001)
    endpoints: []
    tier: large
    context_tokens: 200000
    tool_use_quality: high
    lifecycle: cost-gated
    quota:
      probe_command: null   # wire to a LiteLLM virtual-key budget probe (kb_worker.py BudgetProbe) once a caller opts in
      windows: []
      threshold_pct: null
    gpu_residency: null
    cost_class: metered
    metered: true
    opt_in_required: true

# ── GPU residency policy (§3b) ──────────────────────────────────────────
# "a local model's a(t) is not 1": a(t) = f(VRAM headroom). Never-evict
# models are excluded from eviction math entirely — their VRAM is a fixed
# reservation. Everything else in a co-residency group must fit in what's
# left. preload_check semantics documented here; implemented generically
# in model_registry.py so it reads this data instead of hardcoding numbers.
gpu_residency_policy:
  card: "GTX 1070, 8192 MB (single GPU today; §8 — more GPUs become a placement problem, same policy, more slots)"
  total_vram_mb: 8192
  # Measured 2026-07-21: bge-m3 + gemma3:4b + tei-reranker ~= 6.2/8 GB.
  # Loading something bigger than local-small's footprint on top evicts
  # tei-reranker (LRU-ish ollama/torch behavior) -> silent recall-latency
  # regression. This is the regression the pre-load check exists to catch.
  measured_baseline_mb: 6200
  never_evict_ids: [bge-m3, tei-reranker]
  co_residency_groups:
    interactive-local: [bge-m3, tei-reranker, local-small]
  preload_check:
    description: >
      Before a worker pulls a candidate model onto the GPU it must pass
      this check (see model_registry.py:preload_check): reserve every
      never_evict model's vram_mb unconditionally, subtract whatever else
      is currently resident, and require the candidate's own vram_mb to
      fit in what's left of total_vram_mb. A failing check means "park,
      don't load" — never silently evict a never-evict model.

# ── routing ───────────────────────────────────────────────────────────────
# Tier pools a caller can ask for without naming a model (design §2: "target
# = constraint-set"). metered_opt_in lists the virtual keys that have
# explicitly opted into paid-fallback; empty = no metered model is reachable
# by anyone, satisfying the "no metered API by default" acceptance bar.
routing:
  tiers:
    small: [local-small]
    # paid-fallback listed as a large-tier candidate AFTER codex so resolve()
    # can fail over to it when codex's a(t)=0 (quota parked) — but only for a
    # caller that both passes allow_metered=True AND appears in
    # metered_opt_in below. With metered_opt_in empty (the shipped default)
    # resolve() skips it unconditionally, so it stays unreachable.
    #
    # NB: with the Codex migration there is no quota probe (see the `codex`
    # entry), so a(t) never reads as parked and this failover cannot trigger
    # on quota today.
    large: [codex, paid-fallback]
  metered_opt_in: []   # e.g. ["agent:torgash"] once a human explicitly opts a specific virtual key in